The single most consequential number in Anthropic's S-1 filing is not the revenue growth. It is the percentage of revenue that comes from two unnamed customers.

The prospectus confirms what investors had suspected. Roughly a quarter of Anthropic's 2025 revenue — approximately 4.6 billion dollars — came from two large enterprise clients, each contributing about twelve percent. Either one of them changing course would shift Anthropic's annual revenue by more than a tenth.

The customers are not named in the prospectus. Investors who covered the company earlier in the year have speculated that the two are Cursor and GitHub Copilot. The speculation is plausible. Cursor has publicly stated that its high-end subscription plans are built on long-term partnerships with Anthropic and other frontier model providers. GitHub Copilot integrated Claude in 2024. Neither speculation is confirmed in the filing. Both are consistent with Anthropic's strongest market — AI-assisted software development.

Neither speculation is confirmed in the filing. Both are consistent with Anthropic's strongest market — AI-assisted software development.

The Asymmetry Problem

The concentration is the kind of metric that makes an institutional investor pause. Even if Cursor and Copilot are both customers today, both products offer their users multiple model choices. Cursor lets its users select from Anthropic, OpenAI, Google, and xAI. GitHub Copilot has explicitly moved to a multi-model architecture, with Claude, GPT, and Gemini all available to enterprise customers.

If Claude loses on performance, price, or supply stability for either customer, the customer can switch providers with low migration cost. From Anthropic's perspective, two customers that account for a quarter of revenue are also two customers that could leave with a single quarter's notice.

The risk runs in one direction. Anthropic cannot easily replace those customers. The customers can easily replace Anthropic.

The Cloud Dependency

The second concentration risk sits one layer up. In 2025, Anthropic booked 2.16 billion dollars of revenue through Amazon's and Google's cloud marketplaces. That is forty-seven percent of the company's annual revenue, up from eleven percent in 2023 and thirty-two percent in 2024.

The mechanism is straightforward. AWS, Google Cloud, and Microsoft Azure already have enterprise customers with billing relationships in place. Anthropic's models are available to those customers without a separate procurement cycle. Anthropic, in exchange, gives up a share of every transaction.

The arrangement has been effective. Anthropic's own published data indicates that more than one hundred thousand customers now run Claude on Amazon Bedrock.

The cost is also visible. In 2025, Anthropic paid approximately 351 million dollars in distribution fees to the cloud platforms. Sixteen cents of every dollar booked through a cloud marketplace goes back to the marketplace. By the end of 2025, the company was also carrying 909 million dollars in outstanding customer receivables, sixty percent of which were being collected by the platforms on Anthropic's behalf. That share has been rising.

The Circle That Should Not Work

The cloud platforms are simultaneously Anthropic's investors, its compute suppliers, its distribution channels, and its direct competitors in frontier model development. Each role makes sense in isolation. The combination is what makes the S-1 uncomfortable to read.

The cycle, written out, looks like this. A technology conglomerate invests in Anthropic. Anthropic pays the same conglomerate for compute. The resulting Claude models are sold through the conglomerate's cloud marketplace. The marketplace takes a cut of every sale. The conglomerate uses some of the resulting insight into pricing and demand to inform its own competing model roadmap. The pattern is not unique to Anthropic, but it is unusually concentrated, and it is unusually exposed by the prospectus, which describes the arrangements in plain English rather than burying them in schedules.

OpenAI, which has been more aggressive about structuring its partnerships to preserve optionality, has reportedly complained to investors about Anthropic's accounting. Anthropic books the gross transaction value of cloud-marketplace sales as revenue. OpenAI uses the net value. The difference, on Anthropic's reported numbers, is several billion dollars of upside. Whether the gross or net treatment is correct depends on which accounting standard one applies. Both companies have chosen different ones.

The Five-Hundred-Billion-Dollar Problem

The third concentration sits at the bottom of the capital stack. Anthropic has committed to spending at least 518 billion dollars on AI infrastructure over the next ten years. Eighty percent of the commitments are non-cancellable, on take-or-pay terms that require Anthropic to pay for compute capacity regardless of whether it uses it.

The commitments are concentrated by counterparty. Alphabet's Google is committed to roughly 111 billion dollars of revenue from Anthropic. Amazon is committed to roughly 110 billion. Microsoft is committed to roughly 31 billion. Broadcom-related equipment leases account for roughly 161 billion. The remainder covers xAI, AMD, and other partners.

The Google commitment runs through 2033. The Amazon commitment runs through 2036. The Microsoft commitment is contractually non-cancellable absent material breach by Microsoft.

What the Prospectus Actually Says

The honest reading of the prospectus is that Anthropic is a company with significant revenue, exceptional gross margin pressure, a small number of high-value customers, an even smaller number of high-leverage compute partners, and a long-dated obligation stack that leaves very little margin for strategic missteps.

The optimistic reading is that the same structure that creates concentration also creates alignment. The cloud platforms have made bets that Anthropic succeeds. The two mystery customers have made bets that Anthropic stays competitive. The take-or-pay compute contracts give Anthropic predictable capacity that competitors have to fight for in real time.

Both readings are defensible. The prospectus makes both available. The market will choose between them.

The Real Question

For a public-market investor, the question is not whether Anthropic's technical trajectory is impressive. It clearly is. The question is whether a business with two customers accounting for twenty-five percent of revenue, three counterparties controlling nearly half its distribution, and 518 billion dollars of locked-in infrastructure spend is the kind of business a public shareholder can model with confidence.

The answer depends on how confident one is in the durability of each of those relationships. Cursor and Copilot may be deeply committed to Claude. They may also be one benchmark cycle away from giving the slot to a cheaper model. The cloud partnerships are stable as long as Anthropic remains competitive on the platform's shelf. They become unstable the moment a hyperscaler decides to feature its own model instead.

The prospectus is honest about all of this. That honesty is, in itself, the most useful thing in the document. It tells prospective investors exactly where the risks live. Whether the risks are tolerable is a question each investor will answer for themselves.

The Cycle, in One Sentence

Anthropic takes investor money from companies it sells to. It pays those companies for compute. It sells its products through marketplaces run by those same companies. It commits future cash flows to those companies on take-or-pay terms. And it competes with those companies on the frontier model market.

That sentence describes a real business. It also describes the most concentrated set of dependencies any frontier AI lab has disclosed to date. The IPO will test whether the public market accepts that trade.